LENKIN233/xianyu-monitor-skill v2.0.0-rc.1

on GitHub

This release candidate freezes the documented command/schema surface for final
2.0.0 validation. Guided setup, installation diagnostics, durable monitoring,
AI analysis, and delivery are all included in 2.0; the old 2.1/2.2 roadmap labels
were work packages. This prerelease does not establish live Xianyu, AI-provider,
or notification-service readiness. Stable promotion requires those smokes.
Formal publication requires the clean exact v2.0.0-rc.1 tag.

Breaking

• Direct search and monitor invocations now reject more than 20 pages or 10
  attempts instead of accepting any positive integer. Split larger authorized
  collection into bounded runs; existing persisted tasks remain readable so they
  can be listed, stopped/resumed, or deleted.
• Task commands now report a missing task file as a structured error instead of
  treating it as an empty task list. Create a task first or pass the intended
  absolute --data-file path.

Added

• xianyu demo shows a deterministic synthetic search → validated AI analysis →
  golden evaluation → durable outbox → redacted delivery-preview workflow with
  no credentials, network access, or local writes. Release self-check runs it
  from both the extracted bundle and an empty-HOME copy installation.

• Optional xianyu analyze ranking for sanitized search/monitor JSON with
  no-network preview, explicit external-send consent, strict structured output,
  preview-digest binding, retained recorded-item recovery, and
  OpenAI-compatible/Netlify AI Gateway configuration.

• xianyu state validates an authorized browser-state candidate locally without
  launching a browser or echoing its path or contents.

• Public roadmap, security policy, and release changelog.

• macOS to the Linux/Windows CI matrix.

• Optional xianyu setup orchestration for doctor, visible login handoff, local
  state validation, and a one-page/one-attempt capability test with one final
  path-private JSON result.

• xianyu version machine-readable release/capability discovery and offline
  install --check health states backed by a per-copy hash manifest.

• Deterministic minimal Skill bundles with payload manifest, SPDX 2.3 SBOM,
  external SHA-256, canonical verification, exact dependency lock, clean/tagged
  release gate, and empty-HOME install smoke tests.

• Digest-bound task update previews, credential-free portable task
  export/import, and stopped/no-state import defaults.

• Schema-3 task stores with durable new-item outbox events, stable idempotency
  keys per delivery generation, baseline suppression, listing-field
  allowlisting, explicit ack, and intentional reset/replay support.

• Digest-gated HTTPS delivery adapters for generic webhooks, Bark, and WeCom;
  endpoint secrets stay in environment variables and provider-confirmed events
  are acknowledged one at a time.

• Reproducible AI run hashes and latency evidence, deterministic offline golden
  regression, and atomic private feedback-record capture.

Changed

• Tag CI publishes verified bundles and checksums as GitHub Release attachments
  after downloading and verifying them. Release candidates stay prereleases
  without replacing the latest stable release.

• Added v1 upgrade and rollback guidance, including task-store preservation and
  the distinction between portable definitions and a full recovery backup.

• The 13-command capability surface, contract versions, discovery fields, and
  offline demo envelope are frozen by release-candidate compatibility tests.

• Search and newly created tasks now cap pages at 20 and attempts at 10.

• Item URLs are reconstructed from captured IDs on the canonical Goofish origin;
  remote-provided target links are discarded.

• Cookie-header import rejects duplicate and syntactically invalid Cookie names
  before writing a candidate.

• POSIX task persistence fails closed if user-only file permissions cannot be
  established and verified.

• Task CLI missing-file errors are explicit and recoverable; --data-file works
  on either side of the subcommand.

• Runtime and CI test dependencies are exact-version locked, and GitHub Actions
  are pinned to full commit SHAs.

• Installation health compares copy payloads with the current source even when
  versions match, rejects unknown nested files, and validates linked payload
  completeness before reporting current.

• Guided setup retains child cleanup evidence when the parent is cancelled and
  fails closed as not-established when cleanup cannot be proven.

Security

• Delivery timeouts and cancellation during a send now report possible duplicate
  delivery even when no response was received; pending events remain recoverable.

• Bark/WeCom acknowledgements require integer success codes and reject booleans,
  floating-point values, and strings rather than acknowledging ambiguous replies.

• WeCom text respects its 2048-byte UTF-8 limit while preserving the complete
  item URL and event key. Oversized notification URLs fail before send instead
  of becoming truncated, misleading links.

• Windows state validation compares creation time consistently across stat/fstat
  while retaining exact same-handle change detection, avoiding false state-change
  errors on Python 3.12.

• AI analysis excludes credentials, paths, seller, image, original URL, and
  unknown input fields; API keys are environment-only, redirects are blocked,
  preview consent is bound to the endpoint/request digest, and model
  identity/output are validated before use.

• Delivery endpoints are never accepted in argv or echoed, redirects are
  blocked, response bodies are not exposed, and uncertain sends/acks retain the
  outbox event with explicit duplicate-risk evidence.

• Browser-state preflight validates privacy and content through one open file
  descriptor and rejects mode, owner, identity, or timestamp changes mid-check.

• Prevented API-controlled listing links from becoming clickable arbitrary or
  executable URLs in downstream Agent output.

• Prevented task files from silently falling back to group/world-readable modes
  when fchmod fails.

• Prevented task mutations from silently recreating or overwriting a store that
  was deleted, replaced, or changed after it was loaded.